Local-first explainable AI in C#

Make AI explain itself before anyone has to trust it.

MiniGPTSharp is a compact, client-ready GPT Microscope: a static sales front door, a real ASP.NET Core product, a teaching CLI, JSON exports, reports, tests, and a tiny model people can actually inspect.

View the working repo Watch the static preview

.NET 8 No model API calls CLI + Web + Reports xUnit + CI

Prompt

The capital of France is

Generated trace

capital -> Paris -> model -> next

Next token

capital 21%

France 18%

Paris 16%

Attention

What this sells

A small repo with the posture of a real product.

This is not just a toy model. It is a demonstration that complex AI behavior can be packaged into software that is understandable, testable, local, and beautiful enough to show a client.

For clients

AI you can inspect

Turn abstract model behavior into visible product moments: tokens, probability bars, attention, ablation, reports, and deterministic replay.

For students

Concepts they can touch

Use one tiny, readable model to teach logits, softmax, sampling, prompt context, and why generated text changes.

For engineers

Signals of delivery quality

Clean .NET structure, CLI commands, ASP.NET Core APIs, xUnit tests, CI, docs, scripts, and release-ready positioning.

Brochure versus product

The public site sells it. The repo runs it.

GitHub Pages

Public showcase site

This page is a polished, fast-loading sales surface. It tells the story, previews the interface, and points people to the working repository.

Free static hosting from a public repo
No backend required
Designed for first impressions and portfolio traffic

Downloadable app

Full GPT Microscope

The actual product remains in the repo: an ASP.NET Core app and CLI that run the real C# model locally, with inspect APIs and report export.

Clone the repository
Run the web project on localhost
Use CLI, JSON, reports, tests, and teaching scripts

Interactive static preview

The product story, compressed into one scene.

Toggle the trace below to feel how the local app explains model behavior. The preview is static; the downloadable app runs the same ideas against the C# model.

Model output

Generated continuation

Deterministic argmax

the capital of France is capital Paris model next token learning

Tokenizer

Prompt tokens

5 tokens

Softmax

Next-token probabilities

top 5

Context

Attention heatmap

layer 1

Replay

Generation timeline

first four steps

Proof points

Why it feels bigger than its size.

Explainability as interface

Internals are surfaced as UX, not buried in logs. That is the difference between a demo and a teaching product.

One core, many surfaces

The same C# model powers CLI output, JSON exports, HTML reports, tests, and the ASP.NET Core playground.

Repeatable by design

Seeded generation and local execution make workshops safe, predictable, and easy to explain without cloud dependencies.

Portfolio-grade packaging

CI, docs, roadmap, security notes, scripts, and static hosting make the repo feel release-aware, not improvised.

Illustrated GPT Microscope dashboard with prompt, token, probability, attention, and report panels

Run the real app

Three commands to the full local experience.

The brochure is static. The real value is the repo: build it, generate a report, then run the ASP.NET Core GPT Microscope locally.

dotnet build .\miniGPTCSharp.sln -c Release
powershell -ExecutionPolicy Bypass -File .\scripts\demo-report.ps1
dotnet run -c Release --project .\MiniGPTCSharp.Web\MiniGPTCSharp.Web.csproj --urls http://localhost:5088

Clone on GitHub Read the README